 |
 |
Hi Annika, The effect of age could be considered in three ways: a) The overall age of the subject as a nuisance (e.g., young adults potentially having different measures than elder subjects). b) The difference in time (~11 months) being a nuisance (i.e., subjects with their 2nd visit earlier in time potentially having different measures than those who have their 2nd visit after several months) c) The measured differences between before and after (1st and 2nd visit) being modeled as a function of the time difference between visits (i.e., first visit coded as 0, 2nd visit coded as number of days from the 1st visit, and used as regressor of interest). From your description I believe you want (a), and possibly also (b), but not (c), which could confound the effect of age with the effect of treatment. The exact same considerations apply to the severity degrees, except that these differ according to group, such that you may want to accommodate for the possibility of an interaction. If that is the case then the design would be, for 3 groups (A, B and C): EV1: Code as -1 for visit 1, 1 for visit 2, 0 if not a member of group A EV2: Code as -1 for visit 1, 1 for visit 2, 0 if not a member of group B EV3: Code as -1 for visit 1, 1 for visit 2, 0 if not a member of group C EV4: Age at the respective visit (for fine control, use decimal places, and/or measure age in months instead of years) EV5: Objective severity degree group A EV6: Objective severity degree group B EV7: Objective severity degree group C EV8: Subjective severity degree group A EV9: Subjective severity degree group B EV10 onwards: subject-specific intercepts, that is, one EV per subject coded as 1 for that subject, 0 otherwise. The contrasts are then: C1: 1 -1 0 0 0 0 ... (changes in group A > changes in group B) C2: 1 0 -1 0 0 0 ... (changes in group A > changes in group C) C3: 0 1 -1 0 0 0 ... (changes in group B > changes in group C) There should be one exchangeability block (EB) per subject, and permutations should happen within-block. If you use PALM you can also include permutations of the blocks as a whole (together with permutations within-block). In randomise, just specify the exchangeability blocks file and it will do within-block permutations. That should be it! All the best, Anderson On 13 December 2017 at 04:40, SUBSCRIBE FSL A. Primassin < [log in to unmask]> wrote: > Hello Anderson, > > I am glad that permutation tests are possible. > > Concerning the covariates: > > - Age: > The mean intervall between both measurements is 11.5 months, so age is > changing comparably in all participants (would you consider age then as a > constant covariate and just use age at pre-measurement)? > > - Objective and subjective severities: > Both change in the intervention group, but not significantly in both > control groups. Here I have to specify the following facts: > -The objective disorder severity degree is available for all 3 > participant groups, it is ordinal-scaled (though I have the opportunity to > only include a subscore which is interval-scaled). > - The subjective disorder severity degree is only available for patients > with and patients without intervention (not for healthy participants) and > is ordinal-scaled. > > In general, it would be nice to first calculate the permutations with a > GLM containing no covariates and afterwards adding one ore more to the > model. > > I am very thankful for your advice here! > > All the best, > Annika > >